Application of Deep Learning in Visual Recognition of Unsafe Behaviors of Construction Personnel

Abstract

Unsafe behavior of construction personnel is a primary cause of the frequent safety accidents in the construction industry. The traditional manual inspection method has some disadvantages, such as limited coverage, strong subjectivity, poor real-time performance and so on. With the development of deep learning and computer vision technology, automatic recognition of unsafe behavior based on vision has become an important research direction of construction safety management. This paper systematically reviews the progress in the application of deep learning in visual recognition of unsafe behaviors of construction workers. Firstly, it analyzes the formation mechanism of unsafe behavior and the shortcomings of traditional management methods from the perspective of behavior safety theory. Then according to the technical route, the existing methods are divided into four categories: target detection, pose estimation, time series modeling and lightweight deployment. The principles, representative models and typical application scenarios of various methods are summarized respectively, and the applicability and limitations of various methods in the actual site application are analyzed. The analysis shows that the deep learning recognition method has been widely used in static protective equipment detection, dynamic dangerous action recognition and real-time early warning, but it still faces challenges such as small target detection difficulty, lack of robustness in complex environment, and difficulty in balancing real-time performance and accuracy. Future research should focus on multimodal fusion, video understanding of large models, online continuous learning and interpretable AI, thereby promoting the intelligent upgrading of construction safety management.